1 / 62

EECS 262a Advanced Topics in Computer Systems Lecture 3 Filesystems September 5 th , 2019

EECS 262a Advanced Topics in Computer Systems Lecture 3 Filesystems September 5 th , 2019. John Kubiatowicz Electrical Engineering and Computer Sciences University of California, Berkeley http://www.eecs.berkeley.edu/~kubitron/cs262. Kernel Device Structure. The System Call Interface.

shauna
Télécharger la présentation

EECS 262a Advanced Topics in Computer Systems Lecture 3 Filesystems September 5 th , 2019

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. EECS 262a Advanced Topics in Computer SystemsLecture 3FilesystemsSeptember 5th, 2019 John Kubiatowicz Electrical Engineering and Computer Sciences University of California, Berkeley http://www.eecs.berkeley.edu/~kubitron/cs262

  2. Kernel Device Structure The System Call Interface Process Management MemoryManagement Filesystems DeviceControl Networking Concurrency,multitasking Virtualmemory Files and dirs:the VFS TTYs anddevice access Connectivity Architecture Dependent Code MemoryManager DeviceControl NetworkSubsystem File System Types BlockDevices IF drivers

  3. Today’s Papers • A Fast File System for UNIXMarshall Kirk McKusick, William N. Joy, Samuel J. Leffler and Robert S. Fabry. Appears in ACM Transactions on Computer Systems (TOCS), Vol. 2, No. 3, August 1984, pp 181-197 • Analysis and Evolution of Journaling File SystemsVijayanPrabhakaran, Andrea C. Arpaci-Dusseau, and Remzi H. Arpaci-Dusseau, Appears in Proceedings of the Annual Conference on USENIX Annual Technical Conference (ATEC '05), 2005 • System design paper and system analysis paper • Thoughts?

  4. Head Cylinder Software Queue (Device Driver) Hardware Controller Media Time (Seek+Rot+Xfer) Request Result Review: Magnetic Disk Characteristic Track Sector • Cylinder: all the tracks under the head at a given point on all surface • Read/write data is a three-stage process: • Seek time: position the head/arm over the proper track (into proper cylinder) • Rotational latency: wait for the desired sectorto rotate under the read/write head • Transfer time: transfer a block of bits (sector)under the read-write head • Disk Latency = Queueing Time + Controller time + Seek Time + Rotation Time + Xfer Time • Highest Bandwidth: • Transfer large group of blocks sequentially from one track Platter

  5. Historical Perspective • 1956 IBM Ramac — early 1970s Winchester • Developed for mainframe computers, proprietary interfaces • Steady shrink in form factor: 27 in. to 14 in. • Form factor and capacity drives market more than performance • 1970s developments • 5.25 inch floppy disk formfactor (microcode into mainframe) • Emergence of industry standard disk interfaces • Early 1980s: PCs and first generation workstations • Mid 1980s: Client/server computing • Centralized storage on file server • accelerates disk downsizing: 8 inch to 5.25 • Mass market disk drives become a reality • industry standards: SCSI, IPI, IDE • 5.25 inch to 3.5 inch drives for PCs, End of proprietary interfaces • 1900s: Laptops => 2.5 inch drives • 2000s: Shift to perpendicular recording • 2007: Seagate introduces 1TB drive • 2009: Seagate/WD introduces 2TB drive • 2018: Seagate announces 14TB drives

  6. Disk History Data density Mbit/sq. in. Capacity of Unit Shown Megabytes 1973: 1. 7 Mbit/sq. in 140 MBytes 1979: 7. 7 Mbit/sq. in 2,300 MBytes source: New York Times, 2/23/98, page C3, “Makers of disk drives crowd even mroe data into even smaller spaces”

  7. Disk History 1989: 63 Mbit/sq. in 60,000 MBytes 1997: 1450 Mbit/sq. in 2300 MBytes 1997: 3090 Mbit/sq. in 8100 MBytes source: New York Times, 2/23/98, page C3, “Makers of disk drives crowd even mroe data into even smaller spaces”

  8. Example of Current HDDs • Seagate Exos X14 (2018) • 14 TB hard disk • 8 platters, 16 heads • Helium filled: reduce friction and power • 4.16ms average seek time • 4096 byte physical sectors • 7200 RPMs • 6 Gbps SATA /12Gbps SAS interface • 261MB/s MAX transfer rate • Cache size: 256MB • Price: $615 (< $0.05/GB) • IBM Personal Computer/AT (1986) • 30 MB hard disk • 30-40ms seek time • 0.7-1 MB/s (est.) • Price: $500 ($17K/GB, 340,000x more expensive !!)

  9. Contrarian View • FFS doesn’t matter anymore! • What about Journaling? Is it still relevant?

  10. Storage Performance & Price (2014) 1http://www.fastestssd.com/featured/ssd-rankings-the-fastest-solid-state-drives/ BW: SSD up to x10 than HDD, DRAM > x10 than SSD Price: HDD x30 less than SSD, SSD x4 less than DRAM

  11. Filesystems Background • i-node: structure for per-file metadata (unique per file) • contains: ownership, permissions, timestamps, about 10 data-block pointers • i-nodes form an array, indexed by “i-number” – so each i-node has a unique i-number • Array is explicit for FFS, implicit for LFS (its i-node map is cache ofi-nodes indexed by i-number) • Indirect blocks: • i-node only holds a small number of data block pointers (direct pointers) • For larger files, i-node points to an indirect block containing 1024 4-byte entries in a 4K block • Each indirect block entry points to a data block • Can have multiple levels of indirect blocks for even larger files

  12. A Fast File System for UNIX (4.2 BSD) • Original UNIX FS was simple and elegant, but slow • Could only achieve about 20 KB/sec/arm; ~2% of 1982 disk bandwidth • Problems: • Blocks too small • 512 bytes (matched sector size) • Consecutive blocks of files not close together • Yields random placement for mature file systems • i-nodes far from data • All i-nodes at the beginning of the disk, all data after that • i-nodes of directory not close together • no read-ahead • Useful when sequentially reading large sections of a file

  13. FFS Changes • Aspects of new file system: • 4096 or 8192 byte block size (why not larger?) • large blocks and small fragments • disk divided into cylinder groups • each contains superblock, i-nodes, bitmap of free blocks, usage summary info • Note that i-nodes are now spread across the disk: • Keep i-node near file, i-nodes of a directory together (shared fate) • Cylinder groups ~ 16 cylinders, or 7.5 MB • Cylinder headers spread around so not all on one platter • Two techniques for locality: • Lie – don’t let disk fill up (in any one area) • Paradox: to achieve locality, must spread unrelated things far apart • Note: new file system got 175KB/sec because free list contained sequential blocks (it did generate locality), but an old system has randomly ordered blocks and only got 30 KB/sec (fragmentation)

  14. Skip Sector Track Buffer (Holds complete track) Attack of the Rotational Delay • Problem: Missing blocks due to rotational delay • Issue: Read one block, do processing, and read next block. In meantime, disk has continued turning: missed next block! Need 1 revolution/block! • Solution1: Skip sector positioning (“interleaving”) • Place the blocks from one file on every other block of a track: give time for processing to overlap rotation • Solution2: Read ahead: read next block right after first, even if application hasn’t asked for it yet. • This can be done either by OS (read ahead) • By disk itself (track buffers). Many disk controllers have internal RAM that allows them to read a complete track • Important Aside: Modern disks+controllers do many complex things “under the covers” • Track buffers, elevator algorithms, bad block filtering

  15. Where are inodes stored? • In early UNIX and DOS/Windows’ FAT file system, headers stored in special array in outermost cylinders • Header not stored anywhere near the data blocks. To read a small file, seek to get header, seek back to data. • Fixed size, set when disk is formatted. At formatting time, a fixed number of inodes were created (They were each given a unique number, called an “inumber”)

  16. Where are inodes stored? • Later versions of UNIX moved the header information to be closer to the data blocks • Often, inode for file stored in same “cylinder group” as parent directory of the file (makes an ls of that directory run fast). • Pros: • UNIX BSD 4.2 puts a portion of the file header array on each of many cylinders. For small directories, can fit all data, file headers, etc. in same cylinder  no seeks! • File headers much smaller than whole block (a few hundred bytes), so multiple headers fetched from disk at same time • Reliability: whatever happens to the disk, you can find many of the files (even if directories disconnected) • Part of the Fast File System (FFS) • General optimization to avoid seeksbv

  17. 4.2 BSD Locality: Block Groups • File system volume is divided into a set of block groups • Close set of tracks • Data blocks, metadata, and free space interleaved within block group • Avoid huge seeks between user data and system structure • Put directory and its files in common block group • First-Free allocation of new file blocks • To expand file, first try successive blocks in bitmap, then choose new range of blocks • Few little holes at start, big sequential runs at end of group • Avoids fragmentation • Sequential layout for big files • Important: keep 10% or more free! • Reserve space in the BG

  18. FFS First Fit Block Allocation • Fills in the small holes at the start of block group • Avoids fragmentation, leaves contiguous free space at end

  19. FFS Locality Techniques (Summary) • Goals • Keep directory within a cylinder group, spread out different directories • Allocate runs of blocks within a cylinder group, every once in a while switch to a new cylinder group (jump at 1MB) • Layout policy: global and local • Global policy allocates files & directories to cylinder groups – picks “optimal” next block for block allocation • Local allocation routines handle specific block requests – select from a sequence of alternative if need to

  20. FFS Results • 20-40% of disk bandwidth for large reads/writes • 10-20x original UNIX speeds • Size: 3800 lines of code vs. 2700 in old system • 10% of total disk space unusable (except at 50% performance price) • Could have done more; later versions do

  21. FFS System Interface Enhancements • Really a second mini-paper! • Long file names (14  255 characters) • Advisory file locks (shared or exclusive) • Process id of holder stored with lock => can reclaim the lock if process is no longer around • Symbolic links (contrast to hard links) • Atomic rename capability • The only atomic read-modify-write operation, before this there was none • Disk quotas • Could probably have gotten copy-on-write to work to avoid copying data from user  kernel (would need to copies only for parts that are not page aligned) • Over-allocation would save time; return unused allocation later Advantages: • 1) less overhead for allocation • 2) more likely to get sequential blocks

  22. FFS Summary • 3 key features: • Parameterize FS implementation for the hardware it’s running on • Measurement-driven design decisions • Locality “wins” • Major flaws: • Measurements derived from a single installation • Ignored technology trends • A lesson for the future: don’t ignore underlying hardware characteristics • Contrasting research approaches: improve what you’ve got vs. design something new

  23. Is this a good paper? • What were the authors’ goals? • What about the evaluation / metrics? • Did they convince you that this was a good system /approach? • Were there any red-flags? • What mistakes did they make? • Does the system/approach meet the “Test of Time” challenge? • How would you review this paper today?

  24. BREAK

  25. Transactional File Systems • Better reliability through use of log • All changes are treated as transactions • A transaction is committed once it is written to the log • Data forced to disk for reliability • Process can be accelerated with NVRAM • Although File system may not be updated immediately, data preserved in the log • Difference between “Log Structured” and “Journaled” • In a Log Structured filesystem, data stays in log form • In a Journaledfilesystem, Log used for recovery • Journaling File System • Applies updates to system metadata using transactions (using logs, etc.) • Updates to non-directory files (i.e., user stuff) can be done in place (without logs), full logging optional • Ex: NTFS, Apple HFS+, Linux XFS, JFS, ext3, ext4 • Full Logging File System • All updates to disk are done in transactions

  26. Quick Aside:Log-Structured/Journaling File System • Radically different file system design • Technology motivations: • CPUs outpacing disks: I/O becoming more-and-more of a bottleneck • Large RAM: file caches work well, making most disk traffic writes • Problems with (then) current file systems: • Lots of little writes • Synchronous: wait for disk in too many places – makes it hard to win much from RAIDs, too little concurrency • 5 seeks to create a new file: (rough order) • file i-node (create) • file data • directory entry • file i-node (finalize) • directory i-node (modification time)

  27. LFS Basic Idea • Log all data and metadata with efficient, large, sequential writes • Treat the log as the truth, but keep an index on its contents • Anti-locality for reads! • Great locality for writes (including random writes) • Rely on a large memory to provide fast access through caching • Data layout on disk has “temporal locality” (good for writing), rather than “logical locality” (good for reading) • Why is this a better? Because caching helps reads but not writes! • Two potential problems: • Log retrieval on cache misses • Wrap-around: what happens when end of disk is reached? • No longer any big, empty runs available • How to prevent fragmentation?

  28. LFS Log Retrieval • Keep same basic file structure as UNIX (inode, indirect blocks, data) • Retrieval is just a question of finding a file’s inode • UNIX inodes kept in one or a few big arrays, LFS inodes must float to avoid update-in- place • Solution: an inode map that tells where each inode is (Also keeps other stuff: version number, last access time, free/allocated) • inode map gets written to log like everything else • Map of inode map gets written in special checkpoint location on disk; used in crash recovery

  29. LFS Disk Wrap-Around • Compact live info to open up large runs of free space • Problem: long-lived information gets copied over-and-over • Thread log through free spaces • Problem: disk fragments, causing I/O to become inefficient again • Solution: segmented log • Divide disk into large, fixed-size segments • Do compaction within a segment; thread between segments • When writing, use only clean segments (i.e. no live data) • Occasionally clean segments: read in several, write out live data in compacted form, leaving some fragments free • Try to collect long-lived info into segments that never need to be cleaned • Note there is not free list or bit map (as in FFS), only a list of clean segments

  30. LFS Segment Cleaning • Which segments to clean? • Keep estimate of free space in each segment to help find segments with lowest utilization • Always start by looking for segment with utilization=0, since those are trivial to clean… • If utilization of segments being cleaned is U: • write cost = (total bytes read & written)/(new data written) = 2/(1-U) (unless U is 0) • write cost increases as U increases: U = .9 => cost = 20! • Need a cost of less than 4 to 10; => U of less than .75 to .45 • How to clean a segment? • Segment summary block contains map of the segment • Must list every i-node and file block • For file blocks you need {i-number, block #}

  31. Analysis and Evolution of Journaling File Systems • Write-ahead logging: commit data by writing it to log, synchronously and sequentially • Unlike LFS, then later moved data to its normal (FFS-like) location – this write is called checkpointing and like segment cleaning, it makes room in the (circular) journal • Better for random writes, slightly worse for big sequential writes • All reads go the the fixed location blocks, not the journal, which is only read for crash recovery and checkpointing • Much better than FFS (fsck) for crash recovery (covered below) because it is much faster • Ext3/ReiserFS/Ext4 filesystems are the main ones in Linux

  32. Three modes for a JFS • Writeback mode: • Journal only metadata • Write back data and metadata independently • Metadata may thus have dangling references after a crash (if metadata written before the data with a crash in between) • Ordered mode: • Journal only metadata, but always write data blocks before their referring metadata is journaled • This mode generally makes the most sense and is used by Windows NTFS and IBM’s JFS • Data journaling mode: • Write both data and metadata to the journal • Huge increase in journal traffic; plus have to write most blocks twice, once to the journal and once for checkpointing (why not all?)

  33. JFS Crash Recovery • Load superblock to find the tail/head of the log • Scan log to detect whole committed transactions (they have a commit record) • Replay log entries to bring in-memory data structures up to date • This is called “redo logging” and entries must be “idempotent” • Playback is oldest to newest; tail of the log is the place where checkpointing stopped • How to find the head of the log?

  34. Logging File Systems • Instead of modifying data structures on disk directly, write changes to a journal/log • Intention list: set of changes we intend to make • Log/Journal is append-only • Single commit record commits transaction • Once changes are in the log, it is safe to apply changes to data structures on disk • Recovery can read log to see what changes were intended • Can take our time making the changes • As long as new requests consult the log first • Once changes are copied, safe to remove log • But, … • If the last atomic action is not done … poof … all gone • Basic assumption: • Updates to sectors are atomic and ordered • Not necessarily true unless very careful, but key assumption

  35. Redo Logging • Prepare • Write all changes (in transaction) to log • Commit • Single disk write to make transaction durable • Redo • Copy changes to disk • Garbage collection • Reclaim space in log • Recovery • Read log • Redo any operations for committed transactions • Garbage collect log

  36. Example: Creating a file • Find free data block(s) • Find free inode entry • Find dirent insertion point -------------------------- • Write map (i.e., mark used) • Write inode entry to point to block(s) • Write dirent to point to inode Free Space map … Data blocks Inode table Directory entries

  37. Ex: Creating a file (as a transaction) • Find free data block(s) • Find free inode entry • Find dirent insertion point -------------------------- • Write map (used) • Write inode entry to point to block(s) • Write dirent to point to inode Free Space map … Data blocks Inode table Directory entries tail head pending done start commit Log in non-volatile storage (Flash or on Disk)

  38. ReDo log • After Commit • All access to file system first looks in log • Eventually copy changes to disk Free Space map … Data blocks Inode table Directory entries head tail tail tail tail tail done start commit pending Log in non-volatile storage (Flash)

  39. Crash during logging - Recover • Upon recovery scan the long • Detect transaction start with no commit • Discard log entries • Disk remains unchanged Free Space map … Data blocks Inode table Directory entries tail head pending done start Log in non-volatile storage (Flash or on Disk)

  40. Recovery After Commit • Scan log, find start • Find matching commit • Redo it as usual • Or just let it happen later Free Space map … Data blocks Inode table Directory entries head tail pending done start commit Log in non-volatile storage (Flash or on Disk)

  41. What if had already started writing back the transaction ? • Idempotent – the result does not change if the operation is repeat several times. • Just write them again during recovery

  42. What if the uncommitted transaction was discarded on recovery? • Do it again from scratch • Nothing on disk was changed

  43. What if we crash again during recovery? • Idempotent • Just redo whatever part of the log hasn’t been garbage collected

  44. Redo Logging • Prepare • Write all changes (in transaction) to log • Commit • Single disk write to make transaction durable • Redo • Copy changes to disk • Garbage collection • Reclaim space in log • Recovery • Read log • Redo any operations for committed transactions • Ignore uncommitted ones • Garbage collect log

  45. Can we interleave transactions in the log? head tail • This is a very subtle question • The answer is “if they are serializable” • i.e., would be possible to reorder them in series without violating any dependences • Deep theory around consistency, serializability, and memory models in the OS, Database, and Architecture fields, respectively • A bit more later --- and in the graduate course… start start commit commit pending

  46. Some Fine Points • Can group transactions together: fewer syncs and fewer writes, since hot metadata may changes several times within one transaction • Need to write a commit record, so that you can tell that all of the compound transaction made it to disk • ext3 logs whole metadata blocks (physical logging); JFS and NTFS log logical records instead, which means less journal traffic

  47. Some Fine Points • Head of line blocking: • Compound transactions can link together concurrent streams (e.g., from different apps) and hinder asynchronous apps performance (Figure 6) • This is like having no left turn lane and waiting on the car in front of you to turn left, when you just want to go straight • Distinguish • Between ordering of writes and durability/persistence – careful ordering means that after a crash the file system can be recovered to a consistent past state. • But that state could be far in the past in the case of JFS • 30 seconds behind is more typical for ext3– if you really want something to be durable you must flush the log synchronously

  48. Linux Example: Ext2/3 Disk Layout • Disk divided into block groups • Provides locality • Each group has two block-sized bitmaps (free blocks/inodes) • Block sizes settable at format time: 1K, 2K, 4K, 8K… • Actual Inode structure similar to 4.2BSD • with 12 direct pointers • Ext3: Ext2 w/Journaling • Several degrees of protection with more or less cost • Example: create a file1.datunder /dir1/ in Ext3

  49. Semantic Block-level Analysis (SBA) • Nice idea: interpose special disk driver between the file system and the real disk driver • Pros: simple, captures ALL disk traffic, can use with a black-box filesystem (no source code needed and can even use via VMWare for another OS), can be more insightful than just a performance benchmark • Cons: must have some understanding of the disk layout, which differs for each filesystem, requires a great deal of inference; really only useful for writes • To use well, drive filesystem with smart applications that test certain features of the filesystem (to make the inference easier)

  50. Semantic Trace Playback (STP) • Uses two kinds of interposition: • 1) SBA driver that produces a trace, and • 2) user-level library that fits between the app and the real filesystem • User-level library traces dirty blocks and app calls to fsync • Playback: • Given the two traces, STP generates a timed set of commands to the raw disk device – this sequence can be timed to understand performance implications • Claim: • Faster to modify the trace than to modify the filesystem and simpler and less error-prone than building a simulator • Limited to simple FS changes • Best example usage: • Showing that dynamically switching between ordered mode and data journaling mode actually gets the best overall performance (Use data journaling for random writes)

More Related